Imitation learning is the family of methods that teach an agent a skill from demonstrations of the desired behavior, instead of from a hand-designed reward that the agent must maximize through trial and error. A human teleoperates a robot through a task a number of times, recording the stream of observations and the actions taken; the model then learns to reproduce the mapping from observation to action. For many robotics problems this is far easier than reinforcement learning, because showing a task is often simpler than writing down a reward that exactly captures success.
The simplest form is behavior cloning, which treats the demonstrations as a supervised learning dataset and trains the policy to predict the demonstrator’s action at each step. More sophisticated approaches model the full distribution of good actions - Diffusion Policy, for example, uses a denoising diffusion process so it can represent the several valid ways to perform a task rather than averaging them. A related branch, inverse reinforcement learning, instead tries to recover the reward the demonstrator was implicitly optimizing.
Imitation learning powers much of modern robot manipulation. Systems like Stanford’s Mobile ALOHA collect demonstrations through low-cost teleoperation rigs and learn complex household tasks from as few as 50 demonstrations each, sometimes co-trained with larger shared datasets.
Why business readers should care: imitation learning is what makes it practical to teach a robot a new task by showing it, not programming it - the workflow that companies are betting will let non-experts deploy robots into messy real environments.